1 hour ago
7
The artificial intelligence industry may be building more data centers than it needs while substantial amounts of computing power remain unused, according to David Sherman, AI and financial inclusion strategist at decentralized computing network io.net.
Sherman warned that hyperscalers are increasingly relying on debt to finance massive infrastructure expansion, raising questions about whether those investments will generate sufficient returns.
His comments come as AI stocks face a sharp selloff following hotter than expected US inflation data, which has renewed concerns about borrowing costs and the sustainability of the industry’s spending plans.
“The scale of this selloff should be a warning sign for the AI industry,” Sherman said. He argued that demand for computing power remains strong, but questioned whether continuously building new infrastructure is the most sustainable solution.
AI infrastructure spending faces growing debt concerns
The warning comes as investors reassess the economics of AI infrastructure expansion.
On Thursday, South Korea’s KOSPI index closed 7.57% lower after falling as much as 9.29% intraday. Semiconductor manufacturers Samsung Electronics and SK Hynix declined 9.5% and 11.1%, respectively, according to Reuters.
Meanwhile, financing conditions are becoming more challenging for data center developers.
Reuters reported that investors have grown more selective about funding new projects, particularly those carrying significant debt.
AI infrastructure company Firmus, for example, has faced difficulties supporting a proposed $5 billion IPO amid concerns about its roughly $30 billion debt burden.
Still, investment remains substantial. Debt financing for Asian data center projects has reached $22.3 billion this year, even as equity fundraising declined 4.6% to $5.4 billion.
Research highlights unused AI computing capacity
Sherman’s concerns are supported by recent research into enterprise AI infrastructure.
A September 2026 VentureBeat Intelligence report found that 81% of surveyed enterprises running AI workloads reported GPU utilization of 50% or less, while 31% reported utilization below 25%.
Despite this, 63% planned to expand or replace their AI infrastructure within the next year, while only 10% had a return on investment figure reviewed by their finance teams.
However, the findings were based on a limited enterprise survey and do not establish similar utilization rates among major hyperscalers.
For Sherman, the figures highlight an opportunity to improve existing infrastructure rather than relying exclusively on new data centers.
“The industry has convinced itself that the answer to growing AI demand is simply to build more, when a significant part of the solution could be making better use of what we already have,” he said.
Sherman pointed to distributed computing networks, smarter workload allocation and marketplaces for unused GPU capacity as potential alternatives.
These systems allow companies to access computing resources from multiple providers rather than investing in dedicated infrastructure.
However, distributed computing also faces limitations, including network latency, security requirements and the complexity of coordinating large AI workloads.
Sherman acknowledged that new data centers remain necessary, but argued that better utilization could reduce the industry’s reliance on expensive expansion financed through debt.
The warning comes as investors increasingly scrutinize whether AI infrastructure spending can generate returns sufficient to justify its rising costs.
Disclosure: This article was edited by Estefano Gomez. For more information on how we create and review content, see our Editorial Policy.
English (US) ·